Pruxel PriorAuth Navigator

AI agents that reduce treatment delays by automating prior authorization workflows for healthcare providers.

Comment

Inspiration

Prior authorization is one of the heaviest administrative burdens in US healthcare. Staff spend hours matching a patient's chart against a payer's coverage rules and writing justification letters, and small omissions cause denials that delay care. I wanted to find out whether a team of cooperating AI agents could handle the tedious preparation — so a human reviewer only has to approve, not assemble from scratch.

What it does

Given a clinical note, the requested procedure, and the payer, the system runs three specialized agents:

  1. Clinical Agent reads the free-text note and extracts structured facts — diagnoses, prior treatments and their durations, exam findings — each tied back to where it appeared in the note.
  2. Policy Agent retrieves the relevant payer policy from MongoDB Atlas through the MongoDB MCP server, and turns it into a checklist of criteria.
  3. Decision Agent compares the facts against each criterion, produces a gap analysis, drafts an authorization letter, and returns a recommendation: Ready to Submit, Submit with Caveats, or Not Ready.

A deterministic safety gate (outside the model) guarantees the system never reports "Ready" while a required criterion is unmet or unclear — it fails safe. The app uses synthetic data only and is decision-support with a human in the loop, not for clinical use.

How I built it

  • Google Agent Development Kit (ADK) with a SequentialAgent orchestrating the three sub-agents over shared session state.
  • Gemini 2.5 Flash as the reasoning model.
  • MongoDB Atlas as the policy and case store, queried by the Policy Agent through the MongoDB MCP server — database access is a real MCP tool call, not a hard-coded query.
  • Pydantic output schemas to force structured, validated output from the extraction and decision steps.
  • A Streamlit front end that runs the pipeline and lights up each agent live, then renders the decision, gap analysis, and drafted letter.

Challenges I ran into

  • ADK doesn't allow an agent to use both tools and a strict output schema, so I split the roles: the Policy Agent uses MCP tools, while the Clinical and Decision agents use schemas, and I validate the Policy Agent's JSON downstream.
  • Shutting down the MCP server (a Node process) and asyncio cleanly on Windows was painful. I ended up running the pipeline in a dedicated subprocess and streaming progress back to the UI — which also made the live agent animation reliable.
  • Making the "Ready" verdict trustworthy: I moved the final decision gate out of the model into deterministic Python, so a required gap can never be hand-waved.

What I learned

  • Practical multi-agent orchestration, and where to keep determinism versus LLM judgment.
  • How to wire a real MCP server into an agent so data access is a first-class tool.
  • That the hard part of an agent demo is reliability and honest failure modes, not the happy path.

What's next

  • Atlas Vector Search for semantic policy retrieval across many payers.
  • PDF export of the authorization packet, and "what would flip this to Ready" suggestions.

Built With

Share this project:

Updates